10 m Bin, Weak-Edge Steel Parts: Stitched Views Feed 6D Pose
On a heavy-industry plate cutting line, flat parts drop into scrap bins to be sorted and stacked by specification. The difficulty is the view: a bin plus the pick space above it can span ten metres, and workpieces range from full sheets to small blanks. One camera covering that area spreads its resolution too thin; moving the camera closer loses the whole picture.
A ten-metre bin defeats a single lens twice
Resolution and coverage trade against each other directly. At ten metres, a single sensor divides its pixels across the entire bin, so a weak kerf may occupy less than one pixel. Pulling the lens back to gain detail shrinks the coverage. This is geometry, not a tuning problem, and it is the reason the architecture starts with several cameras rather than a better single one.
Stitching is a calibration job in imaging clothing
The system stitches several camera views into one large panorama. The hard part is not joining pixels but keeping one coordinate system after the join. Each camera has its own pose, so stitching must first solve the relative geometry between cameras and then map pixel positions into a common workpiece frame. If that layer is loose, an accurate pose lands in the wrong place and the robot reaches for a pretty but offset coordinate, which means a miss.
The panorama proposes; the point cloud disposes
Multi-scale object detection runs on the stitched image and a deformable template draws the candidate box. The system then returns to the point cloud to estimate the 6D pose. The split is deliberate: the image answers which part and roughly where, while the point cloud answers how it is oriented with enough precision to grasp. Cut plate has little surface signature, so images struggle to separate near-identical specifications, while point clouds alone struggle to find candidates across ten metres. Each modality works at the scale it handles best.
A rigid template calls normal distortion a wrong part
Cut outlines deviate from the nominal drawing before they ever leave the machine. Thermal distortion, clamping and kerf compensation all shift the contour. A template that only accepts a rigid shape rejects slight deformation, so genuine wrong parts slip through while good parts are stopped repeatedly. Deformable matching allows local stretching of the outline and separates shape change from a different part.
Attention spends itself on kerf and corners
Cross-modal fusion combines point cloud geometry with image texture and contour, and an attention mechanism reweights the matching stage toward kerf lines and corners, the places where near-identical specifications actually differ. Cut edges are faint and surfaces are untextured, so teaching the model where to look beats asking it to look everywhere.
Pose is half the handoff; allocation is the other half
Results go to a multi-robot system that sorts collaboratively and stacks by specification. The hard problem shifts from vision to scheduling: several parts, several robots, several stack positions, so pick order and destination are planned together. Otherwise two robots converge on one part and the stacking order needs correcting later.
Recognition accuracy is above 99 percent and the reference cycle is about 10 seconds, with grasp success and continuous running figures subject to site acceptance results. Sub-millimetre localization across ten metres depends on stitching, multi-scale detection, deformable matching and 6D pose all holding; relax one and the failure shows at the gripper as a miss or an off-centre grab.
What buyers ask about stitching and pose
Is stitching the only way to cover ten metres?
No. A single very high-resolution camera can cover it too. The difference is the pixel budget: the wider the view, the fewer pixels per millimetre, and weak features suffer first. Stitching trades several cameras for coverage and pays with an extra calibration layer.
What does sub-millimetre pose mean for sorting?
The arm can move straight to the computed pose instead of probing and nudging. On stacked or edge-adjacent parts, that difference decides grasp success.
What changes when a new batch has different specifications?
Templates and training samples, mainly. Camera layout and stitching calibration are station assets and normally stay put; only a size range beyond the current view forces a change in camera count or stitching plan.